LFM2.5
Runs on your deviceLiquid AI · May 2026
Liquid AI's small mixture-of-experts - 8B total, 1.5B active per token, so it runs fast on ordinary machines. Strong instruction following and tool use for its size; 10 languages.
- Context
- 125K tokens
- Smallest download
- 5.2 GB
- Runs from
- 12 GB RAM
- License
- LFM Open License v1.0
Will LFM2.5 run on your machine?
This is the same sizing the app uses - not a marketing estimate.
On Apple Silicon, system memory is also graphics memory - pick "None / integrated / Apple Silicon" and read the bars against your Mac's memory. The app checks your real hardware before every download.
Variants & sizes
| Variant | Download | Quantization | Needs | Apple Silicon |
|---|---|---|---|---|
| 8B-A1B (MoE)MoE | 5.2 GB | Q4_K_M | 12 GB RAM | Metal (GGUF) |
The app picks the right variant for your machine and verifies every download. Model files come straight from their official repositories, pinned to exact revisions.
Context window, compared
License
LFM2.5 is released under the LFM Open License v1.0 - read the terms. The weights download to your machine and stay there; Your Own AI adds no accounts, tracking, or lock-in on top.
Frequently asked questions
- Can I run LFM2.5 offline?
- Yes. LFM2.5 is a downloadable model: with Your Own AI it runs entirely on your computer, works with no internet connection, and your conversations never leave your device.
- What hardware does LFM2.5 need?
- The smallest variant (8B-A1B (MoE), a 5.2 GB download) runs on a machine with 12 GB of memory. Larger variants want more memory or a graphics card - the interactive checker on this page grades every variant against your hardware, using the same logic as the app.
- Is LFM2.5 free to use?
- The model weights are released under the LFM Open License v1.0 license and download at no charge. Your Own AI itself is free and open source for local use.
- How much context does LFM2.5 support?
- 125K tokens of context.
Run LFM2.5 in two clicks
Download Your Own AI, pick LFM2.5, and the app sizes it to your hardware. Private by architecture - no account needed for local use.